Google's Gemini 3.5 Flash is based on the Gemini 3 Flash architecture according to its system card, explaining a price increase driven by higher token speeds and interactivity rather than added scale.

> WeirdML oh dear now, it's still well ahead of Chinese models, but it's barely different from the previous Flash. in fact it's pretty much a Gemini Pro 3 to 3.1 update here, they just 3xd the price and added an extra .4. And it's way behind on cost-effectiveness. not RSI-coded.

@dystopi_a no Gemini 3.5 Flash is slightly weaker overall

@scaling01 Do you think it's better than 3.1 Pro?

@scaling01 Knowledge cutoff of January 2025. You can only do so much with RL on an old foundation model. I’m not sure if Google had a failed training run somewhere along the way, but the fact that they are using a base model that is 17 months old is ridiculous.

Oh so this is why they jumped to 3.5.

Its sort of trying to make it look like a big upgrade but its not a new model at all just more post rraining etc.

For a while it felt like these models would just come and come but it seems to be stagnating now and for example with mythos they dont even give it out

Geohotz is maybe right then. All the hype makes us a bit blind.

I think we are in the era of where the models largely dont get better (higher iq), they just suck less (added capabilities in training).

But who knows. Maybe there is a breakthrough on the menu somewhere. But without it, i guess we will face a lot of fake hype going forward

@scaling01 Price hikes never fix performance gaps.

@scaling01 No proof of mythos

@scaling01 it seems like a packaging miss to configure 3.5 flash with the thinking budget and speed they did. They couldve shipped a model half as faster with a bit less thinking that still scores at or better 5.5 low, for cheaper, and the press would be v different

@scaling01 is 3.5 flash the same size as gpt 5.5 and opus?

@scaling01 I found the lagging gemini 3.5 speed much smarter than many models. Then the grok is generally years behind

@scaling01 comparing 3.5flash with opus4.6 is not fair. 3.5Flash is basically unusable.

@scaling01 take a look at agentic scores, that's where it shines

@teortaxesTex 性价比确实拉胯 感觉这波升级诚意不够啊

@teortaxesTex Paying triple for a marginal gain feels like 2021 hype all over again.

@teortaxesTex it is unusable to me. They somehow havent figured out the RL because it doesnt check its work and it just spews tokens.

@scaling01 I think there is a bias in X's opinions. Because people who try AI models in the first 24 hours and care about benchmarks are enthusiast software developers, they give an inflated importance to how good the model is at coding. But well, most people in the world are not coders...

@scaling01 The artificial intelligence for Gemini is always inflated and deflated for Claude which is important to keep in mind. GPT models are quite well defined on Artificial Intelligence.

@scaling01 Actually the latest google update still being slightly behind is hinting at OpenAI and Anthropic having truly started to take off and get velocity benefits on improving performance via their own models.

The intent-missing issue you describe from personal use is often the hardest thing to capture in benchmarks.

SWE-Bench and TerminalBench Hard get at some of it, but a model that technically completes a task while misreading what you actually needed is a failure mode that only shows up in real workflows. That 5-6 month lag estimate on the more demanding benchmarks probably reflects this better than the Artificial Analysis Index does.

@scaling01 Yup it feels like they are slightly behind but they are also focusing on other stuff and capitalizing on personalization via google account data.